Upload CONVERSION_REPORT.md with huggingface_hub

CONVERSION_REPORT.md

@@ -0,0 +1,90 @@
+# ONNX Conversion Report
+
+## Model
+
+- Source: `KBLab/megatron-bert-base-swedish-cased-600k`
+- Architecture: `MegatronBertForMaskedLM` / `MegatronBertModel`
+- Language: Swedish
+- Parameters: about 110M
+
+## Environment
+
+- `conda` env: `onnxconv`
+- `optimum`: `2.1.0`
+- `onnxruntime`: `1.24.4`
+- `transformers`: `4.57.6`
+- `onnx`: `1.21.0`
+
+## Constraint
+
+`optimum.exporters.tasks.TasksManager` in this environment does not support `megatron-bert` ONNX export. Conversion used direct `torch.onnx.export` wrappers instead of `optimum-cli export onnx`.
+
+The default `torch.export`/dynamo ONNX path produced a fill-mask graph that returned all-`NaN` logits in ORT. Final export uses legacy TorchScript ONNX exporter (`dynamo=False`).
+
+## Outputs
+
+### Fill-mask
+
+- Inputs: `input_ids`, `attention_mask`, `token_type_ids`
+- Output: `logits`
+- Shape on validation sample: `[1, 10, 64128]`
+
+### Feature-extraction
+
+- Inputs: `input_ids`, `attention_mask`, `token_type_ids`
+- Outputs: `last_hidden_state`, `cls_embedding`
+- Shapes on validation sample:
+  - `last_hidden_state`: `[1, 18, 768]`
+  - `cls_embedding`: `[1, 768]`
+
+## Validation
+
+### PyTorch vs FP32 ONNX
+
+#### Fill-mask
+
+- Max absolute logit diff: `0.0000688`
+- Mean absolute diff: `0.00000869`
+- Top-5 tokens at mask position matched exactly:
+  - `kung`
+  - `hertig`
+  - `fe`
+  - `bonde`
+  - `konung`
+
+#### Feature-extraction
+
+- `last_hidden_state` max diff: `0.00000593`
+- `last_hidden_state` mean diff: `0.00000108`
+- `cls_embedding` max diff: `0.00000295`
+- `cls_embedding` cosine similarity: `1.0`
+
+### Quantized variants vs FP32 ONNX
+
+#### Fill-mask
+
+| Variant | Status | Notes |
+|---|---|---|
+| fp16 | Good | top-5 exact match, max diff `0.0131` |
+| int8 | Drift | rank-3+ changed |
+| uint8 | Drift | rank-3+ changed |
+| q4 | Moderate drift | rank-3+ changed |
+
+#### Feature-extraction
+
+| Variant | CLS cosine | Mean abs diff | Notes |
+|---|---:|---:|---|
+| fp16 | `0.9999997` | `0.000222` | best compressed variant |
+| int8 | `0.9731` | `0.05597` | notable drift |
+| uint8 | `0.9717` | `0.05750` | notable drift |
+| q4 | `0.7404` | `0.15967` | poor embedding fidelity on this model |
+
+## FP16 repair note
+
+FP16 conversion initially failed in ORT due float32 `Cast` nodes feeding float16 arithmetic in monolithic graphs. Quantization now retargets those casts based on consumer arithmetic patterns before saving, and final fp16 variants load successfully.
+
+## Recommendation
+
+- Default: `fp16`
+- Feature extraction: prefer `fp16`; `int8`/`uint8` beat `q4` on this base model
+- Fill-mask ranking-sensitive workloads can stay on `fp32` if absolute caution needed, but `fp16` matched top-5 exactly on validation sample